Perplexity launched Perplexity Frontier Research on 16 November, an initiative whose stated aim is “an automatic research scientist, an AI system that can conduct research on its own”. The company says the work will be supported by “considerable computational resources, in excess of 10,000 GPUs”.
Chief executive Aravind Srinivas wrote on X that Perplexity is “committing at least 10,000 GPUs to work on multi-agent orchestration and continual learning systems with the goal of building an automated researcher”. The new research site lists 33 earlier works as examples of the agenda, 19 of them arXiv papers, by our count.
Business Pill · AI Agents
An evergreen Business Engineer explainer on AI agents, from answering to doing. It teaches the general idea only and does not discuss Perplexity or its research initiative.
The key insight: As we read it, Perplexity is pairing a stated compute commitment of more than 10,000 GPUs with a promise to publish its models, data and code, and its research page already lists 33 example works, 19 of them arXiv papers, by our count.
What Perplexity Announced
The announcement post, signed by Perplexity Research, names six pillars: “multi-agent orchestration, pre-training, continual learning, reinforcement learning, systems, and evaluations”. Perplexity says it will “release papers, blog posts, data, models, and code”.
The research site, research.perplexity.ai, names Andrew Gordon Wilson as “Research Lead, Perplexity Frontier Research”. It lists roles in San Francisco, Palo Alto, New York and London, with a priority application deadline of December 5, and a research fellowship with predoctoral and postdoctoral tracks.
Perplexity’s own X account said the work will be “released openly, and tested in Perplexity products”. The site puts the test case this way: “our results end up in a product used by millions of curious people, the most challenging test of all”.

What Was Posted on X
Perplexity’s own account announced Perplexity Frontier Research on 16 November at 15:32 UTC, saying the goal is a system that can perform autonomous research, that the work will be supported by more than 10,000 GPUs and that results will be released openly. This piece quotes its line that the work will be tested in Perplexity products.
Chief executive Aravind Srinivas quoted that post six minutes later, writing that Perplexity is committing at least 10,000 GPUs to multi-agent orchestration and continual learning systems with the goal of building an automated researcher. This piece quotes his GPU commitment and his pledge to open source models and tooling.
The 10,000 GPUs, and What Is Stated About Them
The GPU figure appears three ways in Perplexity’s own materials: “in excess of 10,000 GPUs” in the blog post, “more than 10,000 GPUs” on the site and in the company post, and “at least 10,000 GPUs” in Srinivas’s post.
In the four Perplexity sources we read, the chip type, vendor, location, budget and power draw of those GPUs are not given. That is why we do not convert the count into megawatts or dollars here: the answer depends on which accelerator it is.
The site does tie compute to cost. In multi-agent orchestration, it says, “our goal is to provide output that is on the Pareto frontier of quality for any cost the user wishes to spend”, adding “we wish to have the best quality LLM service available at the lowest cost”.
The Track Record on the Research Page
Each of the six research areas on the site carries an “Example works” list. We counted 33 entries: 19 arXiv papers, 13 posts on Perplexity’s blog or research site, and one Hugging Face model page. Pre-training has the most, nine, eight of them on arXiv.
By arXiv identifier, which encodes the month of first submission, 11 of the 19 papers date from 2026, five from 2025, two from 2024 and one from 2022, by our count. The site does not say every listed work was written at Perplexity, and we do not assume it was.
The open releases section lists three open-weight models on Hugging Face, including Decider v1.2 at 27B, and five code and data links, among them SPACE and the WANDR benchmark. Earlier this month we covered Perplexity’s free Decider model.
Update, 16 November 2026: the product side. Srinivas wrote on X at 18:26 UTC that “all Perplexity products are built and dogfooded with Slackbots”. He also claimed that “Perplexity has the best Slack agent (Computer) in the industry right now”, his own assessment, which we have not tested, and said Perplexity “will be migrating that to a persistent Teammate that you can assign any task to”.
His post quoted a Slack case study, which says Computer “started with a few hundred lines of code” and “can connect information across your tools, get other AI models to challenge your thinking, and even kick off a code change”. Forbes ran its story on the launch under the headline “Perplexity’s Gambit To Turn AI Assistants Into Autonomous Workers”; we read only that headline.
Where It Sits in the AI-Scientist Race
Automated science is a crowded lane this month. Today we also covered OpenAI’s new science blog and its GPT-6 physics result and OpenAI’s study of 251 scientists using ChatGPT, and on 14 November Anthropic’s AI scientist hiring push.
Perplexity frames its version around search: the site says research is “a multi-faceted effort, focused on asking the right questions and producing new scientific knowledge as an answer, with search as an integral component”.
The Structural Read
As we read the research site, the goal is framed around search: it describes research as a multi-agent loop that asks questions, gathers what is known, tests and revises, with search as one integral part.
As we read the three wordings of the GPU figure, they agree on the floor, more than 10,000, but none of Perplexity’s materials names the chip, so the count cannot yet be turned into power or dollars.
Aravind Srinivas, Perplexity chief executive, on X, 16 November 2026
“We will publish evals, papers, blogs and open source our models and tooling.”
Three Implications
FOR RESEARCH HIRING The site lists Frontier Research roles in San Francisco, Palo Alto, New York and London, with a priority deadline of December 5.
FOR OPEN MODELS The site already lists three open-weight models on Hugging Face, including Decider v1.2 at 27B, and five code and data links.
FOR COST The site says its orchestration goal is output on the Pareto frontier of quality for any cost the user wishes to spend.
FOR OUTSIDE ACCESS Openness takes other forms elsewhere: today we also covered how OpenAI lets four safety labs inside its models.
The Business Engineer Lens
This story maps onto the Business Engineer framework The Open vs Closed Meta-Framework.
The framework’s core rule: “Close the SCARCE layer (keep proprietary) + Open the ABUNDANT layer (commoditize)”.
As we read it, Perplexity says it will open its models, data and code while its stated test bed is its own product used by millions of people; which of those layers it treats as scarce is our reading, not Perplexity’s. The framework does not discuss Perplexity.

What Is Not Established
Perplexity has not said, in the materials we read, which GPUs make up the 10,000, whether they are owned or rented, or how they are split between this initiative and its products. Those materials give no timeline, budget or headcount target for the team.
The research agenda is a set of goals and open questions, not results. Perplexity was not contacted for this piece.
The Bottom Line
Perplexity is putting more than 10,000 GPUs behind a stated goal of an automated research scientist, with six research pillars and a promise to publish papers, models, data and code. Its research page lists 33 example works, 19 on arXiv, by our count.
95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. We read Perplexity’s 16 November 2026 blog post introducing Perplexity Frontier Research and its research website, research.perplexity.ai, in full, plus Perplexity’s launch post on X and Aravind Srinivas’s post quoting it. The counts of example works and the year split are ours, from the lists on the research site and the arXiv identifiers they link to. We could read only the headline of Forbes’s 16 November article on the launch, so we do not rely on it. Perplexity was not contacted for this piece.
Sources: Perplexity: Introducing Perplexity Frontier Research (16 Nov 2026) · Perplexity Frontier Research website · Perplexity on X (16 Nov 2026) · Aravind Srinivas on X (16 Nov 2026)









